azithromycin/argus-systems_pickup-carrot-no-subversion

Pickup Carrot No Subversion TsFile

This dataset is an Apache TsFile conversion of argus-systems/pickup-carrot-no-subversion, a LeRobot v2.1 robot-manipulation dataset created with LeRobot.

Modalities: Time-series. The converted repository contains numeric robot observations, actions, frame timing, episode/task tags, and mirrored source metadata. Camera videos remain in the original Hugging Face dataset.

Source Dataset

• Original dataset: argus-systems/pickup-carrot-no-subversion
• Repository owner: argus-systems
• License: Apache-2.0
• Robot type: trossen_ai_stationary; Trossen subversion: v1.0
• Task: Pick up the carrot and put it in the bowl.
• Sampling rate: 30 fps
• Source frame files: 58 Parquet episodes, 25,922 rows (episode indexes 1–38, 40–49, and 51–60)
• Source frame layout: data/chunk-000/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-000/{video_key}/episode_{episode_index:06d}.mp4
• Video streams: observation.images.cam_high, cam_low, cam_left_wrist, and cam_right_wrist (58 MP4 files per stream; 232 total)
• The source card has no separate formal author, paper, or completed citation. Its meta/info.json advertises 61 episodes/27,261 frames, while the current revision contains the 58 files and 25,922 rows above.

Converted Files

• TsFile: data/argus_systems_pickup_carrot_no_subversion.tsfile
• Table: argus_systems_pickup_carrot_no_subversion
• Rows: 25,922
• Episodes represented: 58
• Time precision: milliseconds
• meta/ is mirrored from the source; meta/info.json is updated with the TsFile path and conversion record.

Schema

Time is round(timestamp * 1000) milliseconds. The source timestamp column is dropped because it is redundant with Time / 1000 seconds and restarts within each episode.

TAG columns (TsFile table/device dimensions):

• episode_index
• task_index

Scalar FIELD columns:

• frame_index (INT64)
• sample_index (INT64), renamed from source index

Flattened FLOAT FIELD groups:

• action[14] → action_0 … action_13
• observation.state[14] → observation_state_0 … observation_state_13

Conversion Notes

• The shared config-driven lerobot converter was used through the local dataset-specific entry point. All 58 source Parquet files are merged into one table-model TsFile; filter by the TAG columns to select an episode or task.
• Vector columns are flattened to scalar fields, preserving the full source prefix and replacing . with _. No numeric rows or vector dimensions are intentionally dropped.
• Encoding policy: FLOAT/DOUBLE = GORILLA + LZ4; INT32/INT64 and Time = TS_2DIFF + LZ4; BOOLEAN would use RLE + LZ4; TAG values use the TsFile table/device TAG mechanism. The source schema contains no BOOLEAN fields.
• Videos are not duplicated here. They remain in the original repository under videos/. Numeric rows remain frame-aligned through episode_index and frame_index.

Validation

VALIDATION.md and validation_report.json are local artifacts documenting TsFile readability, schema, encoding policy, and row-count equality (25,922 rows). They are intentionally not part of an upload payload.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/argus_systems_pickup_carrot_no_subversion.tsfile")
table_name = "argus_systems_pickup_carrot_no_subversion"
columns = ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"]
with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())